Apple Podcasts  ·  Spotify  ·  Buzzsprout

Paper A

Generative AI Advertising as a Problem of Trustworthy Commercial Intervention

Jingyi Qiu, Qiaozhu Mei — 2026 — arXiv

preprint  ·   ·  read now

Key findings

• Right now, all major AI platforms (Google, Microsoft, OpenAI) keep ads visually separate from AI answers — they show up as labeled boxes above or below the AI response, not mixed into the text itself. But research already shows that when ads are woven into AI responses, people often can't tell the difference between a paid recommendation and a genuine one.
• The authors map out four levels of hidden commercial influence in AI: (1) a product gets mentioned, (2) information is framed to favor a product, (3) the AI nudges you toward a specific action (like clicking 'buy'), and (4) over time, the AI subtly shapes what you like and prefer. The higher the level, the harder it is to detect or disclose.
• AI can influence your choices in ways that never look like an 'ad' at all — by choosing which sources to cite, which products to compare, or which tool to use next. One study cited found that just changing how a product is described to an AI shopping agent was enough to shift which product 'won' in recommendations.
• Organic ChatGPT referrals to e-commerce sites are already converting at higher rates and generating more revenue per visit than paid social ads — meaning AI is already a serious commercial channel even without explicit ads.

Marketing implications

• If you run AI-assisted campaigns or use ChatGPT/Copilot for content, know that your audience likely can't tell when AI-generated text has been commercially shaped — which means trust is fragile and transparency is your competitive edge. Label AI-assisted content voluntarily before regulators force you to.
• Start paying attention to 'generative engine optimization' (GEO) — getting your brand cited by AI answers organically is already driving better conversion rates than paid social. That's where to invest time now.
• If you're building or buying AI ad tools, the safest and most durable products will be ones that can show what commercial influence was applied and why — auditability will become a selling point as scrutiny increases.

Paper B

Traditional statistical representations outperform generative AI in identifying expert peer reviewers

Vicente Amado Olivo, Tereza Jerabkova, Jakub Klencki, John Carpenter et al. — 2026 — arXiv

preprint  ·   ·  watchlist

Key findings

• A simple statistical method called TF-IDF — which just counts how often specific words appear in a document — correctly identified a true expert in the top 25 recommendations 79.5% of the time. GPT-4o mini only managed 51.5% — nearly 30 percentage points worse.
• GPT-4o mini and similar AI language models struggled because they blend together the meanings of words too broadly. In highly specialized fields, tiny differences in vocabulary (e.g., specific sub-topic names) matter enormously for finding the right expert — and AI smooths those differences away.
• Traditional statistical approaches were also cheaper and more transparent. LLMs require far more computing power and their reasoning is harder to audit, yet they performed worse on this specialized task.
• The study validated results two ways: using who authored which proposal as a proxy for expertise, and using reviewers' own self-reported expertise labels. Both approaches confirmed the same ranking of methods.

Marketing implications

• If your team is building or buying an AI tool to match content, products, or experts to highly specialized audiences (e.g., B2B technical buyers, niche professional communities), test whether a simple keyword-frequency approach outperforms the AI before paying for the AI — it might, especially in narrow domains.
• For any AI-powered recommendation or matching system you use, ask the vendor: 'What's your benchmark accuracy compared to a simple keyword-matching baseline?' If they can't answer, that's a red flag.
• When matching content to audiences in specialized niches (technical SaaS, medical, legal, financial), preserve specific jargon rather than letting AI paraphrase it. Exact vocabulary often carries more signal than 'smarter' semantic interpretations.

Apple Podcasts  ·  Spotify  ·  Buzzsprout